Runtime Memory Management in Many-core Systems

OF THE DISSERTATION Runtime Memory Management in Many-core Systems By Hossein Tajik Doctor of Philosophy in Computer Science University of California, Irvine, 2016 Professor Nikil Dutt, Chair With the number of cores on a chip continuing to increase, we are moving towards an era where many-core platforms will soon be ubiquitous. Efficient use of tens to hundreds of cores on a chip and their memory resources comes with unique challenges. Some of these major challenges include: 1) Data Coherency – the need for coherency protocol and its induced overhead poses a major obstacle for scalability of many-core platforms. 2) Memory requirement variation – concurrently running applications on a many-core platform have variable and different memory requirements, not only across different applications, but also within a single application; in this dynamic scenario, static analysis may not suffice to capture dynamic behaviors. 3) Scalability – inefficiency of a central management makes distributed management a necessity for many-core platforms. To address all these issues, this dissertation proposes a comprehensive approach to manage available memory resources in many-core platforms equipped with Software Programmable Memories (SPMs). The main contributions of this dissertation are: 1) We introduce SPMPool: a scalable platform for sharing Software Programmable Memories. The SPMPool approach exploits underutilized memory resources by dynamically sharing SPM resources between applications running on different cores and adapts to the overall memory requirements of multiple applications that are concurrently executing on the many-core platform.